docs/interfaces.md

Benchmark interfaces

The three interfaces represent the same broad behavioral tasks with different user-context representations.

Direct

Direct provides behavioral history in input_text.

• Belief: prior market questions, choices, and position-derived confidence, followed by a target market.
• Trade: prior current-market and other-market trades plus the target decision context.

Belief Direct has train, validation, and test files. Trade Direct also has all three splits.

Profile

Profile replaces or summarizes substantial behavior history with a generated wallet profile.

Belief profile rows add:

profile_generator_model, generated_profile

Trade profile rows add:

profile_generator_model, wallet_profile

The profile files identify the generator model at row level and store the generated representation alongside each model-ready instance.

Belief Profile has train, validation, and test files. Trade Profile releases validation and test benchmark files plus profiles_test.jsonl.

Retrieval

Retrieval augments the prediction context with records from support wallets excluded from the fixed experiment-wallet set.

Belief retrieval rows add:

retrieval_method, retrieved_support

The accompanying metadata specifies exact event-slug matching, strict historical filtering, same-wallet exclusion, and top-5 latest matches.

Trade retrieval rows add:

retrieval_method, retrieved_market_traces

Trade Retrieval releases validation and test files. Individual rows expose the retrieval method and retrieved traces.

Cross-interface alignment

Use sample_id to compare interfaces. For Belief, released Direct, Profile, and Retrieval split counts match.